Comparative Evaluation of Deep Traffic Sign Classification Models Under Visual Degradations and Interpretability Analysis

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Comments

1- The topic of traffic sign classification models is one of the important issues in intelligent transportation systems and autonomous vehicles and has gained great importance in recent years. On the other hand, the interpretability of models plays an important role in improving their effectiveness among engineers. Therefore, the importance of this issue and its compliance with the objectives of the special issue "Visual Features in Computer Vision Applications" is confirmed. However, due to the importance of this issue in recent years, there are relatively many articles in this field that question the innovation of these designs. Therefore, to prove the innovation, a detailed table is practically necessary that carefully reviews and analyzes the related works and shows why a new research paper in this field is needed and what special prominence the present paper has compared to previous papers. This issue is very necessary.

2- Reliable detection of warning signs is very important. Two approaches in this field should be considered.

* Statistical approach based on probability theory and proving the probabilistic power of the calibrated model for estimation.

*Simulation approach and presentation of results based on the criteria for evaluating deep learning models and using statistical tests to test the validity of the results and the significance of the advantages.

This reviewer can only agree to accept this article if both approaches are pursued simultaneously and their results confirm the effectiveness of the present article compared to previous articles.

3- Paying attention to the compatibility of deep learning models for recognizing sign images in degraded visual conditions is very important and beautiful. In this regard, data enhancement methods and noise removal methods are available. For example, there are successful attempts to remove foggy or dusty noise. Environmental simulators also exist in this field. The respected authors should conduct appropriate laboratory studies and tests in this regard and present the results separately.

4- Theoretical analyses are required to prove computational efficiency, while the article has neglected this issue.

5- To prove interpretability, the important Grad-CAM method has been considered. However, it is recommended that new Grad-CAM developments be considered and their results presented and compared in different sections of the article.

6- The basic, residual, light, and attention enhancement classifiers in an integrated framework are of great interest to the scientific community. However, putting these sections together is an engineering model for solving practical problems and not an innovative approach for specialized tasks. Recently, knowledge distillation learning models with strong theoretical and practical support have been proposed. It is recommended to use these approaches to improve the innovative level of the article.

7- Does adding a convolutional block attention module (CBAM) to a standard convolutional neural network (CNN) model, a ResNet18 model, a standard MobileNetV2 model, and MobileNetV2 seek to solve challenges or make improvements? Please define these challenges quantitatively in the form of a problem statement, and then show, based on computational experiments, how much improvement has been achieved for any challenge. In particular, a detailed study is conducted according to practical scenarios of traffic sign recognition under visual problem conditions. The results can be analyzed and classified.

8- The German dataset is excellent for presenting the results. However, it is not clear how strong the model is on other similar datasets. To prove domain adaptation, it is recommended to consider at least one other dataset to evaluate the model results.

9- Comparison of clean classification and robustness testing in blur, occlusion, and low light conditions is a very prominent point of the article. However, the main question is whether this is due to image augmentation methods or related to the innovation of the convolutional block module? Where has this issue been reviewed and analyzed?

10- Analysis of computational efficiency by separating training time and inference time and comparing these with related works is strongly recommended.

11- According to the results of the article, ResNet18 has the strongest classification performance on the clean set, and MobileNetV2 remains a competitive compact alternative. However, this is not a conclusion of the present paper and is a proven issue in the deep network literature. It is expected that the obvious results will not be reproduced and the focus will be on new and potentially valuable results for the current work and other future works.

12- On the other hand, it is expected that the models resistant to visual degradation conditions will be discussed separately, and based on the results of the models, the practical solutions can be extracted for each type of degradation. Figure 8 is not suitable for this purpose and does not help effectively.

13- Minor commons:

13-1- The presence of Figure 1 in a scientific paper does not help at all.

13-2- Gap Analysis is purely descriptive and not analytical

13-3- Column 1 (year) and column 4 (dataset) of Table 1 are unnecessary. The name of the dataset can be in the table title.

13-4- The architecture presented in Figure 2 does not correspond to the nature of the methods used. For example, the interpretability of the output shown in Figure 2 is similar to that of tabular methods and not the Grad-CAM method. A major revision of this figure is necessary.

13-5-Figure 8 and Table 5 reflect the same information, and one can be omitted.

13-6-Figure 10 does not show useful results from Grad-CAM. Because in all three subfigures, the shape itself is highlighted, and no help is seen from the content inside the sign to distinguish symptoms.

13-7-Figures 11 and 12 beautifully illustrate the problems of Grad-CAM. What is the solution? At least practical suggestions should be provided.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

1. Section 1.4: I don't see any novelty as authors claim; instead, I can say you refer to it as application-level contributions. The second novelty, as you mentioned, is actually the generalization of the model to any real data. and so on.
2. Irregular fonts in tables.
3. if fig 1 as a whole or parts are AI generated, authors should mention. I will suggest that such workflow diagrams can be simplified just by block diagram. The caption is important which needs to be very detailed.
4. why only one dataset is used? How to quantify genralization? 
5. and only 10 epoch. Are the authors sure that the models converged properly?
6. you have tested model on blur data. what was the blur kernal? you can quantify your model approach towards the severity of blur or noise in images.
7. did the authors do any real quantification of interpretability improvement?

I will honestly recommend authors to review their work/experimentation and rewrite it before it is thoroughly reviewed. Right now novelty is weak, work/manuscript has major flaws. I mentioned some above. I believe authors can do much better. 

Comments on the Quality of English Language

extensive revsion is required.
many typos throughout

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

Summary

The manuscirpt serves as a baseline comparision paper between different ML approaches for the Traffic sign classification problem. Its novelty lies in the new evaluation architecture and harsh condition testing. The comparison itself is useful for the general readers who want to know the performance ofdiffernet models and further develop these methods.

The paper is well organized and the overal presentation is very clear. There remains several clarity problems, I would recommend a minor revision.

1. Section 4.2

I do not observe a clear improvement between MobileNetV2 and MobileNetV2+CBAM in terms of validation accuracy shown in Figure 6. In Figure 7, MobileNetV2+CBAM appears to achieve a validation accuracy below 0.97, whereas MobileNetV2 achieves a value slightly higher than 0.97. These results seem to suggest that CBAM has little effect, or may even slightly reduce the accuracy.

Therefore, I am not fully convinced by the authors’ statements that “the evidence observed in the results for the combined performance and visualization supports the usefulness of CBAM with this classification setting” and that “this strongly supports the use of CBAM as an interpretable refinement.”

Please clarify this point.

2 Blur, Occlusion, and Low Light

I may have missed this point, but I could not find a clear description of how the “blur,” “occlusion,” and “low light” conditions were introduced. Did the authors manually classify the raw data into these three categories, or were artificial modifications applied to the original dataset to generate these conditions?

If artificial treatments were applied, please clarify the corresponding data treatment protocol.

3. Figure-2 True Data?

Are the results shown in Figure 2 (Workflow – Part 4) obtained from real test data? The results do not appear to be consistent with those reported later in Table 3.

If these are real results, please clarify in the manuscript how the data were obtained and why they differ from the results shown in Table 3, which seems to represent a similar test.

If the figure is intended only as a schematic illustration, I recommend using real test results instead of artificial/example data, as the current presentation may lead to confusion for readers.

4. Line-473

It seemed to be typo. It should be 182 e-6 instead of 182-e7. Please double check.

Author Response

Please see the attachment.

Author Response File: Author Response.docx

Reviewer 4 Report

Comments and Suggestions for Authors

Title: Comparative Evaluation of Deep Traffic Sign Classification Models under Visual Degradations and Interpretability Analysis

This study presents a reliability-oriented comparison of Baseline CNN, ResNet18, MobileNetV2, MobileNetV2 enhanced with the Convolutional Block Attention Module, and knowledge-distilled MobileNetV2. All models were trained independently using five random seeds on a fixed stratified split of the German Traffic Sign Recognition Benchmark. The work is good in its field; however, it currently contains many major and minor points (shown below), which should be carefully considered and corrected.

Remarks to the Authors: Please see the full comments.

1-In Abstract, it is essential to give clear emphasis to the methodology of this comparison, whether the assessment measures are fair, and whether the results are reproducible among the five selected models: baseline CNN, ResNet18, MobileNetV2, MobileNetV2 enhanced with the Convolutional Block Attention Module, and knowledge-distilled MobileNetV2.

In addition, various performance metrics can be used to evaluate a model, such as accuracy, speed, reliability, and robustness to degradations, computation cost, and many other. What metrics are used to evaluate the performance of the selected models? Please state them clearly.

Finally, it is recommended to include some specific quantitative calculation results to highlight the best models compared to other relevant models.

2- Did this work compared to other published related works? What are the differences between it and them? Besides, what specific methodological comparison (key feature of the methodology) did the authors consider in their work?

3-This study states that “this study presents a reliability-oriented comparative framework for traffic sign classification”. Therefore, it is recommended to clarify the importance of this framework and its basic concepts.

4-The manuscript can be further improved by revising its grammar and deleting the repeated information. For example:

-At line 118, the word “the” must be removes from the sentence “To address this problem the we present a specific evaluation”.

- The manuscript repeatedly uses specific terms such as "accurate, robust, statistically reliable, calibrated, transferable, efficient, and interpretable." This can be avoided in various ways by using a different approach.

-In lines 843-844, it is stated that “The experiments provide a more balanced interpretation of CBAM than the original manuscript.”. What the authors mean by “manuscript”?

- The paper needs more organization, and the structure of the paper can be added for more clarity.

5-Based on lines 151-154, how was the transferable property valued based on the contributions given?

6- A section for related works (different from the literature review) should be added, including works that have compared some models in the same field, then summarized them, and provided a brief and balanced overview of a broad field in a table that distinguishes the current work from those other existing works (Table 1 provided a good comparison, but it is not the required comparison, in addition, why was reference [21] included?).

7- Expand the models’ architecture description by including filter sizes, number of filters, learning rate, dropout values, number of pooling, and any other implementation details for reproducibility.

8- It is recommended to add a descriptive block diagram that shows the workflow in a concise and clear way instead of the diagram shown in Figure 1 (as Figure 1 is well organized but confuses the reader and presents many details close to an AI diagram).

9- Five model configurations were evaluated in the study. However, there are other models such as VGG-16, LeNet-5, TrafficSignNet, EfficientNetB7 and different algorithms (many versions of Yolo). What are the reasons for choosing only these five?

10- For McNemar paired significance tests, provide further explanation on how the data are structured?

11- The visual presentation of Figure 4 should be enhanced.

12- There is a lot of repeated information in different sections that could be combined or organized in another way, such as including subsection 4.8 in section 5, and the subsections of section 5 could be made more concise or summarized in a table in a clear way. Please check the entire manuscript for this issue carefully.

Comments on the Quality of English Language

The English could be improved to more clearly express the research.

Author Response

Comment 1: In Abstract, it is essential to give clear emphasis to the methodology of this comparison, whether the assessment measures are fair, and whether the results are reproducible among the five selected models: baseline CNN, ResNet18, MobileNetV2, MobileNetV2 enhanced with the Convolutional Block Attention Module, and knowledge-distilled MobileNetV2.

In addition, various performance metrics can be used to evaluate a model, such as accuracy, speed, reliability, and robustness to degradations, computation cost, and many other. What metrics are used to evaluate the performance of the selected models? Please state them clearly.

Finally, it is recommended to include some specific quantitative calculation results to highlight the best models compared to other relevant models.

Response: The revised abstract and Results section now report specific numerical results. ResNet18 achieved the strongest five-seed mean clean performance, with 0.9856 ± 0.0093 accuracy and 0.9817 ± 0.0134 macro-F1. It also achieved the best GTSDB macro-F1 of 0.9389 in the representative Seed-42 external evaluation.

Location: Abstract; Section 4.1; Section 4.5.

Comment 2: Did this work compared to other published related works? What are the differences between it and them? Besides, what specific methodological comparison (key feature of the methodology) did the authors consider in their work?

Response: The revised Literature Review now explicitly compares the present study with direct model-comparison studies and reliability-oriented studies. Table 1 and Appendix Tables A1–A2 show that prior works typically cover only subsets of the proposed evaluation dimensions, whereas the present study jointly evaluates clean performance, five-seed stability, bootstrap intervals, McNemar tests, calibration, robustness, external GTSDB transfer, efficiency, CBAM/KD effects, and Grad-CAM concentration.

Location: Literature Review; Table 1; Appendix Tables A1–A2.

Comment 3: This study states that “this study presents a reliability-oriented comparative framework for traffic sign classification”. Therefore, it is recommended to clarify the importance of this framework and its basic concepts.

Response: We clarified the framework in the Introduction and Methodology. Reliability-oriented evaluation is now defined as joint assessment of clean predictive performance, training stability, paired statistical evidence, probabilistic calibration, robustness under controlled visual corruptions, external generalization, computational efficiency, and Grad-CAM-based diagnostic behavior.

Location: Section 1.1; Methodology opening.

Comment 4: The manuscript can be further improved by revising its grammar and deleting the repeated information. For example:

-At line 118, the word “the” must be removes from the sentence “To address this problem the we present a specific evaluation”.

- The manuscript repeatedly uses specific terms such as "accurate, robust, statistically reliable, calibrated, transferable, efficient, and interpretable." This can be avoided in various ways by using a different approach.

-In lines 843-844, it is stated that “The experiments provide a more balanced interpretation of CBAM than the original manuscript.”. What the authors mean by “manuscript”?

- The paper needs more organization, and the structure of the paper can be added for more clarity.

Response: The manuscript was edited throughout to improve grammar and organization. Repetitive wording was reduced, and sections were reorganized to move from motivation and related work to methodology, results, discussion, limitations, and conclusion. We also removed revision-history wording such as “original manuscript.”

The affected sentence was removed/replaced during the revision of the Introduction. The revised Introduction no longer contains this grammatical error.

Location: Full manuscript, especially Introduction, Methodology, Results, and Discussion.

Comment 5: Based on lines 151-154, how was the transferable property valued based on the contributions given?

Response: Transferability is now explicitly evaluated through external validation on 360 cropped GTSDB signs mapped to compatible GTSRB labels. The manuscript reports GTSDB accuracy, GTSDB macro-F1, clean GTSRB macro-F1, and macro-F1 domain drop, and also discusses the limitation that this external set is small and class-imbalanced.

Location: Section 3.1; Section 4.5; Limitations.

Comment 6: A section for related works (different from the literature review) should be added, including works that have compared some models in the same field, then summarized them, and provided a brief and balanced overview of a broad field in a table that distinguishes the current work from those other existing works (Table 1 provided a good comparison, but it is not the required comparison, in addition, why was reference [21] included?).

Response: We revised Section 2 to include direct comparative traffic sign recognition studies. These include comparisons of CNN/ResNet/MobileNet, compact BNN architectures, pretrained CNNs with LIME, and adverse-condition CNN comparisons with Grad-CAM. A condensed comparison is given in Table 1, and the full matrix is provided in Appendix Tables A1 and A2.

Location: Section 2; Table 1; Appendix Tables A1–A2.

Comment 7: Expand the models’ architecture description by including filter sizes, number of filters, learning rate, dropout values, number of pooling, and any other implementation details for reproducibility.

Response: We added Table 2 for architecture details and retained Table 4 for training configuration. Table 2 includes filter sizes, pooling, dropout, classifier heads, CBAM details, distillation setup, and parameter counts. Table 4 reports input size, crop size, batch size, optimizer, learning rate, weight decay, epochs, early stopping, pretrained weight status, mixed precision, and knowledge-distillation role.

Location: Section 3.3, Table 2; Section 3.9, Table 4.

Comment 8: It is recommended to add a descriptive block diagram that shows the workflow in a concise and clear way instead of the diagram shown in Figure 1 (as Figure 1 is well organized but confuses the reader and presents many details close to an AI diagram).

Response: We revised Figure 1 to present the workflow more clearly. It now summarizes data context, model system, reliability evaluation, and results/insights. We retained limited icons and visual markers for readability, but simplified and reorganized the figure so that it functions as a clearer workflow diagram. It was inspired from AI but completely drawn from scratch in Powerpoint.

Location: Figure 1, after Methodology opening.

Comment 9: Five model configurations were evaluated in the study. However, there are other models such as VGG-16, LeNet-5, TrafficSignNet, EfficientNetB7 and different algorithms (many versions of Yolo). What are the reasons for choosing only these five?

Explain why VGG-16, LeNet-5, TrafficSignNet, EfficientNetB7, and YOLO were not selected.

Response: The model-architecture section now explains that the selected five models represent targeted comparison categories rather than an exhaustive architecture survey. It also explains that YOLO-family models are detection architectures and would require a different scene-level detection-and-classification protocol.

Location: Section 3.3 Model Architectures.

Comment 10: For McNemar paired significance tests, provide further explanation on how the data are structured?

Response: We expanded the statistical testing subsection to describe the paired prediction structure, the four paired outcomes, the discordant counts, and the null hypothesis. Table 6 also defines (n_{01}) and (n_{10}).

Location: Section 3.4; Table 6.

Comment 11: The visual presentation of Figure 4 should be enhanced.

Response: Figure 4 and its caption were revised. The caption now states the software used to generate the figure and the inference-measurement protocol. The figure is interpreted together with Table 11.

Location: Section 4.6, Figure 4 and Table 11.

Comment 12: There is a lot of repeated information in different sections that could be combined or organized in another way, such as including subsection 4.8 in section 5, and the subsections of section 5 could be made more concise or summarized in a table in a clear way. Please check the entire manuscript for this issue carefully.

Response: We removed the repeated summary-style subsection and revised the Discussion to interpret the findings directly by evaluation dimension. The Results section now ends with Grad-CAM diagnostics before moving into the Discussion.

Location: End of Section 4.7 and beginning of Section 5

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Thanks for the modifications. The font size of some tables, especially Table 1, is very small and they are not readable.

Author Response

Comment1: Thanks for the modifications. The font size of some tables, especially Table 1, is very small and they are not readable.

Response: 

Thank you for highlighting the readability issue. We have revised the presentation of the literature-review tables and standardized the table typography throughout the manuscript. Most importantly, the original dense literature-review matrix has been replaced in the main manuscript by a substantially condensed and more readable Table 1, containing two representative studies from each of the four literature-review themes. The number of columns and entries was reduced, the descriptive text was shortened, and the comparative tick/cross columns were retained to preserve the intended visual comparison. The revised table now fits clearly within a standard portrait page without requiring excessive font reduction. The corresponding explanatory paragraph was also revised to clarify the study-selection strategy and to direct readers to the complete literature matrix.

Location of changes in the revised manuscript:

• Section 2.4, “Attention, Interpretability, and Reliability-Aware Evaluation,” page 7: the paragraph immediately preceding Table 1 now explains that two representative studies were retained from each literature-review theme and that the complete matrix is provided as supplementary material.
• Table 1, page 8: revised and condensed table entitled “Condensed comparison of representative traffic sign studies. Two representative studies are retained from each literature-review theme.”
• Appendix Table A1, page 29: the complete literature-review matrix has been relocated to a dedicated sideways landscape page so that all columns and entries remain available without overcrowding the main manuscript.

We also reviewed the remaining tables to maintain consistent font sizing, spacing, decimal formatting, model naming, and caption style across the manuscript.

Reviewer 2 Report

Comments and Suggestions for Authors

Comparative Evaluation of Deep Traffic Sign Classification Models under Visual Degradations and Interpretability Analysis

The paper presents a reliability-oriented comparative evaluation of five deep learning architectures for traffic sign classification, such as BaselineCNN, ResNet18, MobileNetV2, MobileNetV2+CBAM, and knowledge-distilled MobileNetV2. The evaluation spans clean-set performance, statistical significance, calibration, severity-wise robustness under four degradation types, external generalization (GTSDB), computational efficiency, and Grad-CAM interpretability. The framework is the study's primary contribution rather than any new architecture. 
The following are several comments that need to be addressed.

1. First of all, you must revise the abstract. Keep it within the limit of 200 words.
2. The critical concern is single-seed experimentation, which is the limitation of the study. Your central claim is reliability and statistical rigor; reporting results from a single training run substantially weakens the conclusions. Bootstrap CIs capture test-set sampling uncertainty but say nothing about training variance. The ResNet18 vs. MobileNetV2+CBAM McNemar result (p=0.7493) could plausibly reverse under a different seed. The authors acknowledge this point in the Limitations section but should ideally provide at least 3-5 seed runs with mean ± std for the main metrics or explicitly quantify training stability.
3. All the tables (all) should be in a uniform font. Table 1 is unreadable. You can move table 1 to supplementary data. or just keep the most important columns. Now it is overly dense.
4. Revise the figures too. Figure 5 is unclear and uninformative. All four shown Grad-CAM examples depict the same class (True=17, Pred=17 correct predictions) for all four models. This makes it impossible to visually assess failure cases, misclassifications, or cross-model differences in attention behavior, which is precisely what the authors claim to analyse. The qualitative heatmap set should include: (a) a correct prediction, (b) a misclassification, and (c) a failure case for each model, enabling comparative visual inspection.
5. 360 cropped samples are a very small external test set, especially for 43 classes. Several classes may have 0-1 samples, making macro-F1 unstable. The authors should report class coverage in the GTSDB split (how many of the 43 classes are represented?) and ideally report per-class or grouped results to avoid misleading macro averages dominated by underrepresented classes.
6. The central rectangle occlusion removes the exact center of the sign, which often contains the most discriminative content (digits, arrows, symbols). This is a reasonable experimental choice but should be explicitly justified and compared against random occlusion. As written it may overestimate occlusion sensitivity.
7. The temperature T=4.0 and α=0.35 are given without justification or ablation. Were these tuned, or taken from a prior work? Given that KD is one of the paper's five models, readers may expect some rationale for these choices. Even a brief sensitivity note would suffice.
8. The calibration analysis shows MobileNetV2-KD is worse than standard MobileNetV2 in ECE/NLL/Brier. The paper correctly notes this but does not explore why. Soft-label distillation with T=4.0 should in principle, improve calibration (smoother targets). A brief discussion of possible explanations (e.g. underfitting, temperature not optimal) would strengthen this section.
9. The Grad-CAM center-concentration score is introduced as a quantitative proxy but no baseline or validation is offered. Does a higher concentration score actually correlate with better accuracy? Does it correlate with human expert judgements of explanation quality?  If you include a scatter plot of concentration score vs. prediction confidence, it would be informative.

I appreciate your efforts in revising the manuscript, but it still has enough room for improvement. I believe you can do it. Focus on the quality of presentation. 

Comments on the Quality of English Language

The authors still need extensive revision to improve the quality of the presentation. The current manuscript contains numerous mistakes, awkward sentences, etc. 

Author Response

Reviewer Comment

The paper presents a reliability-oriented comparative evaluation of five deep learning architectures for traffic sign classification, such as BaselineCNN, ResNet18, MobileNetV2, MobileNetV2+CBAM, and knowledge-distilled MobileNetV2. The evaluation spans clean-set performance, statistical significance, calibration, severity-wise robustness under four degradation types, external generalization (GTSDB), computational efficiency, and Grad-CAM interpretability. The framework is the study's primary contribution rather than any new architecture.

Comment 1: First of all, you must revise the abstract. Keep it within the limit of 200 words.

Response: Thank you for this comment. The abstract has been completely rewritten and shortened to remain within the requested 200-word limit. The revised abstract now clearly states the five-model comparison, the five-seed experimental protocol, the reliability dimensions considered, the principal findings, and the evaluation-level contribution of the study. Detailed secondary results and excessive methodological descriptions were removed to improve clarity and concision.

Location of change:

• Abstract, page 1, lines 1--18.

Comment 2: The critical concern is single-seed experimentation, which is the limitation of the study. Your central claim is reliability and statistical rigor; reporting results from a single training run substantially weakens the conclusions. Bootstrap CIs capture test-set sampling uncertainty but say nothing about training variance. The ResNet18 vs. MobileNetV2+CBAM McNemar result could plausibly reverse under a different seed. The authors should ideally provide at least 3--5 seed runs with mean ± std for the main metrics or explicitly quantify training stability.

Response: We agree that bootstrap resampling of one test prediction set does not quantify variability caused by stochastic model training. We therefore repeated the complete principal training and evaluation protocol using five predetermined random seeds: 42, 123, 777, 2024, and 3407. The same fixed stratified train, validation, and test assignments were used in all runs so that variability primarily reflects stochastic initialization, minibatch ordering, augmentation, and optimization rather than changes in test composition. The principal clean-set results are now reported as mean ± sample standard deviation across the five independently trained checkpoints. Seed 42 is retained only as the representative checkpoint for detailed qualitative figures, reliability diagrams, confusion matrices, and Grad-CAM cases. The manuscript now explicitly distinguishes:

1. training variability, quantified across five independent seeds; and

2. test-sample uncertainty, quantified through bootstrap confidence intervals for a fixed checkpoint.

The revised results show that model ranking does vary across seeds and that the apparent advantage of one strong architecture over another is not always stable. Consequently, the manuscript no longer interprets a single McNemar result as definitive evidence of general model superiority.

Location of changes:

• Introduction, page 3: final paragraph of the Introduction, explaining five-seed evaluation and the distinction between seed variability and bootstrap uncertainty.

• Section 1.4, Novelty and Contributions, pages 4--5: first and second contribution statements.

• Section 3.2, Reproducibility and Multi-Seed Training Protocol: description of the five seeds, fixed split, reporting protocol, bootstrap interpretation, and paired-testing strategy.

• Section 4.1, Clean GTSRB Classification Performance and Statistical Reliability, page 15.

• Table 4, page 15: clean performance reported as mean ± sample standard deviation across five independent seeds.

• Section 5.8, Limitations and Future Directions, page 26: acknowledgement that five runs still provide limited power for seed-level non-parametric tests.

Table 4 now reports, for example:

• ResNet18 accuracy: (0.9856 \pm 0.0093);

• MobileNetV2 accuracy: (0.9813 \pm 0.0057);

• MobileNetV2+CBAM accuracy: (0.9801 \pm 0.0107); and

• MobileNetV2-KD accuracy: (0.9743 \pm 0.0088).

Comment 3: All the tables should be in a uniform font. Table 1 is unreadable. You can move Table 1 to supplementary data or keep only the most important columns. It is overly dense.

Response: The literature-review presentation has been substantially reorganized. The original dense comparison matrix was removed from the main narrative and replaced by a condensed Table 1 containing two representative studies from each literature-review theme. The number of rows and the amount of descriptive text were reduced while retaining the comparative coverage columns for clean performance, efficiency, robustness, external evaluation, attention, calibration/statistics, and explainability. The complete literature-review matrix has been moved to a dedicated sideways appendix table so that the full information remains available without reducing the main-text font to an unreadable size. Table typography, spacing, decimal precision, model naming, and caption style were also standardized throughout the manuscript.

Location of changes:

• Section 2.4, page 7: paragraph immediately preceding Table 1, explaining the study-selection strategy.

• Table 1, page 8: condensed literature comparison with two representative studies per theme.

• Appendix Table A1, pages 28--30: complete literature-review matrix presented in landscape orientation.

Comment 4: Figure 5 is unclear and uninformative. All four shown Grad-CAM examples depict the same class and correct predictions. The qualitative heatmap set should include a correct prediction, a misclassification, and a failure case for each model.

Response: Figure 5 has been completely replaced. The revised figure now presents three diagnostic scenarios for each of the four advanced models: a correct prediction, a misclassification, and a high-confidence failure case. The original image, true label, predicted label, confidence, and Grad-CAM activation are displayed to support direct comparison across ResNet18, MobileNetV2, MobileNetV2+CBAM, and MobileNetV2-KD. The figure caption and associated discussion were revised to clarify that Grad-CAM is used as a diagnostic visualization rather than as proof of causal model reasoning.

Location of changes:

• Section 3.13, Grad-CAM Diagnostic Analysis, page 14: revised methodological description.

• Figure 5, page 20: new correct, misclassified, and failure-case heatmaps.

• Table 11, pages 20--21: quantitative Grad-CAM diagnostic results.

• Section 4.7, pages 20--22: revised interpretation of qualitative and quantitative explanation behaviour.

• Section 5.6, page 25: expanded discussion of Grad-CAM strengths and limitations.

The revised Figure 5 caption explicitly states that each model is shown under the three requested diagnostic scenarios.

Comment 5: The 360 cropped samples are a very small external test set, especially for 43 classes. Several classes may have 0-1 samples, making macro-F1 unstable. The authors should report class coverage and ideally per-class or grouped results.

Response: We agree that reporting only a global macro-F1 value would be potentially misleading for this external subset. The revised manuscript now explicitly reports that the GTSDB evaluation contains 360 cropped signs representing 38 of the 43 GTSRB-compatible classes. Five classes are absent, and several represented classes contain only one or two samples. The revised external analysis therefore reports accuracy and weighted-F1 alongside macro-F1, together with class support, absent classes, low-support classes, and per-class performance. The interpretation has also been moderated: the experiment is described as limited cross-dataset validation on a class-compatible subset rather than comprehensive external validation of all 43 classes. The external YOLO label indices were harmonized with the GTSRB class taxonomy before evaluation, and no GTSDB samples were used for training, validation, calibration, or hyperparameter selection.

Location of changes:

• Section 1.4, Novelty and Contributions, page 4: external-validation contribution.

• Section 3.11, External Generalization on GTSDB, page 13: class coverage, support limitations, label harmonization, and interpretation.

• Section 4.5, External Generalization on GTSDB, page 18.

• Figure 3, page 18: external macro-F1 comparison.

• Table 9, page 18: external accuracy, macro-F1, clean GTSRB macro-F1, and domain drop.

• Section 5.8, page 26: explicit limitation concerning the 360-sample, 38-class subset.

The revised manuscript now states that the external set covers 38 of 43 classes, with five classes absent.

Comment 6: The central rectangle occlusion removes the exact center of the sign, which often contains the most discriminative content. This is reasonable but should be explicitly justified and compared against random occlusion.

Response: The central occlusion protocol is now explicitly defined as a targeted worst-case corruption rather than an estimate of average natural occlusion. The manuscript explains that the center of a sign frequently contains the principal digit, symbol, or directional arrow, and central masking is therefore expected to be especially damaging. To contextualize this worst-case condition, we added a support-matched random-occlusion analysis. At each severity level, a rectangle with the same area as the corresponding central mask was placed at repeated deterministic random locations. Performance was averaged across placements. This enables separation of sensitivity to targeted removal of the sign interior from sensitivity to untargeted obstruction.

Location of changes:

• Research Question 3, page 4: central versus support-matched random occlusion.

• Section 1.4, pages 4--5: robustness contribution.

• Robustness methodology, page 12: explicit justification of central occlusion and description of the random-occlusion comparison.

• Table 3, page 13: severity definitions for the controlled corruption protocol.

• Section 5.8, page 26: limitation explaining that neither synthetic protocol fully represents naturally occurring obstruction.

Comment 7: The temperature (T=4.0) and (\alpha=0.35) are given without justification or ablation. Were these tuned or taken from prior work? A brief sensitivity note would suffice.

Response: The knowledge-distillation methodology has been expanded to explain both parameters and to distinguish the prespecified main setting from the auxiliary sensitivity analysis. The main five-seed experiments use (T=4) and (\alpha=0.35), corresponding to a moderate temperature and a comparatively larger contribution from the teacher's soft targets. These values were fixed before the final five-seed evaluation and were not selected using the held-out test set. A focused Seed-42 sensitivity analysis was added using:

[
T \in {2,4,6}, \quad \alpha=0.35,
]

and:

[
\alpha \in {0.35,0.50,0.70}, \quad T=4.
]

The manuscript explicitly states that this is a local sensitivity analysis and does not establish a globally optimal KD configuration. The results were not used retrospectively to alter the prespecified five-seed model.

Location of changes:

• Section 3.4, Knowledge-Distilled MobileNetV2, page 11: KD loss equation, parameter definitions, rationale, and sensitivity design.

• Section 1.4, page 4: contribution statement referring to the focused sensitivity analysis.

• Section 5.8, page 26: limitation that only one teacher--student architecture and one prespecified main configuration were evaluated.

Comment 8: The calibration analysis shows MobileNetV2-KD is worse than standard MobileNetV2 in ECE, NLL, and Brier score. Soft-label distillation should in principle improve calibration. A brief discussion of possible explanations would strengthen this section.

Response: The calibration interpretation has been revised using the five-seed evidence. The new results show that knowledge distillation does not produce a consistent calibration advantage: it improves some calibration measures in some seed runs but degrades them in others. Accordingly, the manuscript no longer presents KD as uniformly better or worse calibrated. The discussion now explains that smoother teacher targets do not guarantee calibrated student probabilities. Possible reasons include the selected temperature and hard/soft loss weighting, early stopping based on predictive performance rather than calibration, student underfitting, transfer of teacher overconfidence, and seed-dependent checkpoint selection. The local sensitivity experiment also shows that calibration and classification performance vary with (T) and (\alpha).

Location of changes:

• Section 3.4, page 11: KD formulation and sensitivity analysis.

• Section 4.3, Calibration Analysis: revised calibration results and comparison between standard and distilled MobileNetV2.

• Section 5, Discussion: expanded interpretation of seed-dependent KD calibration.

• Section 5.8, page 26: explicit statement that distillation did not provide a consistent calibration advantage across seeds.

Comment 9: The Grad-CAM center-concentration score is introduced as a quantitative proxy, but no baseline or validation is offered. Does higher concentration correlate with accuracy or human expert judgement? A scatter plot of concentration versus confidence would be informative.

Response: We agree that center concentration should not be interpreted as a validated measure of explanation quality. The revised methodology now defines it only as a spatial localization descriptor. A higher score indicates that more Grad-CAM activation lies within the predefined central region, but it does not establish greater faithfulness, correctness, or human interpretability.

We added three analyses:

1. Pearson correlation between center concentration and prediction confidence;

2. Spearman rank correlation between center concentration and prediction confidence; and

3. comparison of mean concentration for correct and incorrect predictions.

A new scatter plot was also added to visualize center concentration against prediction confidence for the representative Seed-42 checkpoints. The observed relationships are weak, inconsistent, and model-dependent. In some cases, incorrect predictions exhibit concentration equal to or greater than correct predictions. Therefore, the score is retained only as a descriptive diagnostic measure. No human-expert ratings were collected, and this is now explicitly acknowledged as a limitation.

Location of changes:

• Research Question 6, page 4.

• Section 3.13, Grad-CAM Diagnostic Analysis, page 14: definition and limitations of the concentration score.

• Figure 5, page 20: correct, misclassified, and failure-case Grad-CAM examples.

• Table 11, pages 20--21: mean concentration, standard deviation, Pearson correlation, Spearman correlation, and correct-versus-incorrect comparison.

• Figure 6, page 21: scatter plot of center concentration versus prediction confidence.

• Section 4.7, pages 20--22: revised interpretation.

• Section 5.8, page 26: acknowledgement that no human-expert or perturbation-based validation was performed.

Comment 10: The manuscript still has room for improvement. Focus on the quality of presentation.

Response:
The manuscript has undergone a broad presentation revision. The abstract, research questions, contribution statements, methodology, results interpretation, limitations, table formatting, figure captions, and model terminology were revised for consistency. Dense content was moved out of the main narrative, representative visualizations were replaced, multi-seed results were promoted to the main results tables, and limitations are now reported more explicitly.

Specific presentation improvements include:

• a shortened abstract within 200 words;

• a condensed and readable Table 1;

• relocation of the complete literature matrix to a sideways appendix table;

• standardized model names and numerical formatting;

• replacement of the original Grad-CAM figure;

• addition of the concentration-confidence scatter plot;

• explicit separation of multi-seed training variability from bootstrap uncertainty; and

• expanded captions and cross-references for tables and figures.

Locations of major presentation changes:

• Abstract, page 1.

• Research Questions and Contributions, pages 4--5.

• Table 1, page 8.

• Methodology, pages 9--14.

• Table 4, page 15.

• Table 9 and Figure 3, page 18.

• Figure 5 and Table 11, page 20.

• Figure 6, page 21.

• Limitations, page 26.

• Appendix Table A1, pages 28--30.

We thank the reviewer for encouraging a more rigorous and readable presentation. The revised manuscript now treats reliability as a multi-dimensional and repeated-run evaluation problem rather than as a comparison based primarily on one clean-set accuracy result.

Reviewer 4 Report

Comments and Suggestions for Authors

Most of the comments have been addressed and no more are needed. It is only recommended to review the presentation of the Figures in the results section.

Comments on the Quality of English Language

The English could be more improved to more clearly express the research.

Author Response

Reviewer Comment:
Most of the comments have been addressed and no more are needed. It is only recommended to review the presentation of the Figures in the results section.

Response:
Thank you for this helpful recommendation. We carefully reviewed the presentation of all figures in the Results section and revised Figures 2–7 to improve their readability, consistency, and publication quality. In particular, we regenerated the robustness, external generalization, accuracy–efficiency, and Grad-CAM-related figures with clearer formatting, larger visual elements, improved axis labels, more readable legends, and consistent model naming. Figures 5 and 6 were also enlarged and reformatted to make the Grad-CAM examples easier to inspect. Figure 7 was redrawn with improved marker visibility to reduce overplotting and better show the relationship between Grad-CAM center concentration and prediction confidence. The figure captions were also refined where necessary to better describe the revised visual content. No numerical results, experimental settings, or scientific interpretations were changed.

Changes Made:
Figures 2–7 in the Results section have been revised for clearer presentation and improved readability. The captions of the revised figures have also been updated accordingly.

Round 3

Reviewer 2 Report

Comments and Suggestions for Authors

1. Abstract: unnecessary first sentence. high accuracy? clean benchmark? very unclear terms used. Abstract must contain numerical results.
2. "Evaluation covered clean-set performance and training stability, bootstrap confidence intervals, paired statistical testing, probabilistic calibration, severity-wise robustness under blur, central and random occlusion, low-light, and Gaussian-noise corruptions, external generalization to 360 cropped GTSDB signs, computational efficiency, and Grad-CAM-based diagnostic analysis." This is a long, meaningless sentence with no numerical results.
3. Abstract: What do you mean by stronger architecture? 
4. GTSRB doesn't make sense as a keyword. Keywords usually are methodological and technological terms. Similar for explainable artificial intelligence.
5. Introduction: Until citation [1]. You could easily summarize this into a single sentence on the importance of traffic sign recognition.
6. The introduction, including 1.1-1.4. Huge repetition. You can have this introduction on a single page. Keep it concise and focused.
7. Section 2: Literature review: Paragraph 1. Isn't it all already explained in the introduction? 
8. Section 2 is also unnecessarily extended. Including sections 2.1-2.4. You could provide a better numerical evaluation of these studies. 
9. Methodology: The framework includes dataset preparation..... Isn't it already an available dataset? You claimed it in the data availability statement.
10. Line 305-309. Line 314-318. What is this gobbledygook? Did any of the authors read it before including any sentences from AI? 
11. Table 1. Is it produced through AI? There is no declaration. What is Clean Perf. 
12. Lines 35-37, Lines 321-325? Why this much repetition? And more.
13. Figure 1. What did you add in results and insight? Follow any standard paper based on deep learning for object recognition. You can learn how to present your work properly.
14. Section 3.4? Where are the model architectures? 
15. Figure 4. Which software and how many images did you use to obtain these results? You should have at least arranged the legends in the figure before including it in the manuscript.
16. Table 8. it could be rearranged 
17. Figure 5. unreadable
18. Table for abbreviations is unnecessary. Where did you use RQ in the manuscript? 
19. I highly doubt the authors have put any effort to revise or even read it before submission.
20. Very unethical. References 1 to 7. I can understand the researchers' urge for citations. But can you explain the citation from 3-7? Works like food recognition, tomato detection ......... does it make any sense in work like traffic sign recognition? Just type "traffic sign" on Scholar; you can have thousands of works to cite. I really doubt if you have done any literature review before planning this manuscript and drawing the research gap. 

Author Response

Comment 1: Abstract: unnecessary first sentence. High accuracy? Clean benchmark? Very unclear terms used. Abstract must contain numerical results.

Response: Thank you for this comment. We rewrote the abstract to remove vague and unsupported phrasing such as “high accuracy” and unclear references to “clean benchmark” performance. The revised abstract now begins with a concise importance statement, clearly identifies the five evaluated models, states the five-seed protocol, and reports numerical results. Specifically, it now includes the mean accuracy and macro-F1 of ResNet18, MobileNetV2, and BaselineCNN, the expected calibration error of ResNet18, the GTSDB macro-F1 in the representative Seed-42 external evaluation, and the effect of severe central occlusion.

Comment 2: “Evaluation covered clean-set performance and training stability, bootstrap confidence intervals, paired statistical testing, probabilistic calibration, severity-wise robustness under blur, central and random occlusion, low-light, and Gaussian-noise corruptions, external generalization to 360 cropped GTSDB signs, computational efficiency, and Grad-CAM-based diagnostic analysis.” This is a long, meaningless sentence with no numerical results.

Response: We revised this part of the abstract. The evaluation dimensions are now stated more clearly and are followed immediately by quantitative findings. The abstract no longer relies on a long list of evaluation items without results; instead, it reports specific performance, calibration, external-validation, and robustness values.

Comment 3: Abstract: What do you mean by stronger architecture?

Response: We avoided using “stronger architecture” as a vague standalone term. The revised manuscript now names the exact evaluated architectures and interprets model strength only through reported empirical results across clean performance, calibration, robustness, external transfer, computational efficiency, and Grad-CAM diagnostics.

Comment 4: GTSRB does not make sense as a keyword. Keywords are usually methodological and technological terms. Similar for explainable artificial intelligence.

Response: We revised the keywords to focus on methodological and technological terms rather than dataset names or overly broad labels. The revised keywords are: traffic sign classification; convolutional neural networks; robustness evaluation; visual corruptions; probabilistic calibration; knowledge distillation; attention mechanism; Grad-CAM.

Comment 5: Introduction: Until citation [1]. You could easily summarize this into a single sentence on the importance of traffic sign recognition.

Response: We condensed the opening of the Introduction. It now briefly states the importance of traffic sign recognition for intelligent transportation systems, advanced driver-assistance systems, and autonomous-driving pipelines, and then moves directly to the real-world problem of visual variability.

Comment 6: The Introduction, including 1.1–1.4, has huge repetition. You can have this introduction on a single page. Keep it concise and focused.

Response: We substantially condensed and reorganized the Introduction. Repeated statements about traffic sign recognition, benchmark accuracy, robustness, and reliability were removed or merged. The revised Introduction now focuses on the problem motivation, dataset use, reliability-oriented gap, selected five-model comparison, novelty/contributions, research questions, and paper structure.

Comment 7: Section 2: Literature review: Paragraph 1. Is it not all already explained in the Introduction?

Response: We revised the opening paragraph of the Literature Review so that it no longer repeats the Introduction. It now starts with direct model-comparison studies in traffic sign recognition and explains how the present study extends these works through multi-seed stability, paired statistical testing, calibration, severity-wise corruption robustness, external GTSDB transfer, computational efficiency, and Grad-CAM concentration.

Comment 8: Section 2 is also unnecessarily extended, including Sections 2.1–2.4. You could provide a better numerical evaluation of these studies.

Response: We revised Section 2 to improve focus and comparison. We added direct traffic sign model-comparison studies and included numerical values where available. We also added a condensed comparison table in the main manuscript and moved the complete comparison matrix to the appendix, where it is divided into two tables for readability.

Comment 9: Methodology: The framework includes dataset preparation. Is it not already an available dataset? You claimed it in the data availability statement.

Response: We clarified this point. The revised manuscript now explicitly states that no new traffic sign dataset was created. GTSRB was used for training, validation, clean testing, and controlled corruption testing, while cropped and label-compatible GTSDB signs were used only for external validation. The wording was revised to avoid implying that the authors created or prepared a new dataset.

Comment 10: Line 305–309. Line 314–318. What is this gobbledygook? Did any of the authors read it before including any sentences from AI?

Response: We acknowledge that the earlier wording was unclear and unsuitable. The methodology opening was rewritten in direct scientific language. The revised text now clearly explains dataset use, fixed partitioning, shared preprocessing and training settings, five random seeds, and the role of Seed 42 for representative qualitative figures. The manuscript was reread and edited to remove unclear or AI-like phrasing.

Comment 11: Table 1. Is it produced through AI? There is no declaration. What is Clean Perf.

Response: We revised Table 1 for clarity and changed “Clean Perf.” to “Clean performance.” The table now compares studies across clearly defined dimensions, including efficiency, robustness, external evaluation, attention, calibration/statistical reliability, and XAI. We have also added a declaration of AI-assisted use, clarifying that AI assistance was used only for drafting, restructuring, and formatting selected tables and schematic figures, while all scientific content, numerical results, interpretations, and final table entries were checked and verified by the authors.

Comment 12: Lines 35–37, Lines 321–325? Why this much repetition? And more.

Response: We removed and condensed repeated statements across the Introduction, Methodology, Results, and Discussion. The revised manuscript avoids repeatedly restating the same reliability dimensions and instead uses a structured evaluation table to summarize the metrics once.

Comment 13: Figure 1. What did you add in results and insight? Follow any standard paper based on deep learning for object recognition. You can learn how to present your work properly.

Response: We revised and simplified Figure 1 to present the experimental workflow more clearly. The revised figure links data context, model system, reliability evaluation, and results/insights. It was reorganized to better match the structure of a deep-learning experimental workflow, while retaining limited visual markers for readability.

Comment 14: Section 3.4? Where are the model architectures?

Response: We added a dedicated model architecture table. The new table reports implementation details for BaselineCNN, ResNet18, MobileNetV2, MobileNetV2+CBAM, and MobileNetV2-KD, including convolutional layers, kernel sizes, pooling, dropout, classifier heads, CBAM structure, distillation setup, and parameter counts.

Comment 15: Figure 4. Which software and how many images did you use to obtain these results? You should have at least arranged the legends in the figure before including it in the manuscript.

Response: We revised Figure 4 and its caption. The caption now states that the figure was generated using Python and Matplotlib from saved evaluation outputs. It also clarifies that inference time was measured in the same execution environment as Table 11, with models in evaluation mode and gradient computation disabled. The figure legend and layout were also revised for clearer presentation.

Comment 16: Table 8. It could be rearranged.

Response: We rearranged the CBAM ablation table. The revised version now compares MobileNetV2 without CBAM, MobileNetV2 with CBAM, and the absolute difference under the same no-augmentation setting. This makes the effect of CBAM easier to interpret.

Comment 17: Figure 5 unreadable.

Response: We revised the Grad-CAM presentation. The original dense figure was split into two separate figures: one for correctly classified examples and one for misclassifications and high-confidence failure cases. We also replaced abbreviations with full labels such as “True Class,” “Predicted Class,” “Prediction Confidence,” and “Grad-CAM center concentration.”

Comment 18: Table for abbreviations is unnecessary. Where did you use RQ in the manuscript?

Response: The abbreviation table was removed. The research questions are retained to guide the evaluation structure, and the Results section now follows the corresponding evaluation dimensions: clean performance and statistical reliability, calibration, robustness, CBAM/KD effects, external generalization, computational efficiency, and Grad-CAM diagnostics.

Comment 19: I highly doubt the authors have put any effort to revise or even read it before submission.

Response: We regret that the earlier version gave this impression. The manuscript has now been carefully revised throughout. We rewrote the abstract, condensed the Introduction, revised the literature review, corrected the citations, added direct traffic sign comparison studies, clarified the methodology, added architecture details, revised the figures and tables, added reproducibility information, and improved the interpretation of results. The revised manuscript was reread and checked by the authors before resubmission.

Comment 20: Very unethical. References 1 to 7. I can understand the researchers’ urge for citations. But can you explain the citation from 3–7? Works like food recognition, tomato detection, etc. do not make sense in traffic sign recognition. I doubt whether a literature review was done properly.

Response: We acknowledge the concern and revised the reference list substantially. The revised manuscript now cites traffic-sign-specific benchmark and related-work references, including GTSRB, GTSDB, and recent traffic sign recognition, detection, robustness, efficiency, external-generalization, and interpretability studies. We also added direct comparative traffic sign studies to the Literature Review and comparison tables to strengthen the research gap and positioning.